The AI agent that attacked 20 companies in Japan
Artex, the open-source Chinese agent that orchestrates Claude, GPT and DeepSeek, was used against some 20 Japanese firms. State attribution still lacks proof.
The investigation by Japanese security firm GMO, published by Kyodo on October 20, 2026, describes an offensive that bears little resemblance to conventional intrusion campaigns. The tool is Artex, an open-source autonomous agent created by Chinese engineer Li Puhua, capable of orchestrating models such as Claude, Generative Pre-trained Transformer (GPT) and DeepSeek to map networks, craft lures and carry out intrusions with little human oversight. Six dedicated attack units targeted Japan in campaigns segmented by country and sector, according to the GMO investigation published by Kyodo on October 20, 2026.
The scale is striking. Around 80 companies were hit in total, according to Kyodo, with South Korean banks among the targets outside Japan. Jiji Press confirms six servers with Artex installed and eight Japanese firms with data leaks between September and October 2026. Bloomberg details about 20 targets in Japan, including Daiwa, East Japan Railway Company (JR East) and SoftBank, with statements from Daiki Fukumori, a former Interpol officer now at GMO. The economic daily Maeil Business Newspaper (MK), via the Wall Street Journal, traces the origin: more than 20 IP (Internet Protocol) addresses in over 10 countries operated the open-source agent created by Li Puhua.
The Japanese government reacted the same day. According to the Mainichi, the cybersecurity minister declared that Tokyo is treating the case as malicious use of artificial intelligence. The wording matters: the cabinet did not attribute the operation to Beijing, only acknowledged that autonomous agents have entered the inventory of threats the state needs to regulate and combat.
From manual intrusion to agent operations
The technical difference explains the alarm. Classic intrusion tools require operators who choose targets, write lure emails and escalate privileges by hand. Artex automates the chain: it scans for vulnerabilities, adapts the lure to the victim’s sector and runs attempts in parallel, with the large language models doing the writing and decision-making work that once occupied entire teams. Six servers dedicated to Japan point to industrialization, with campaigns segmented by country and sector, according to Jiji Press.
Cyber offensives have become agent operations, and the human operator now supervises swarms.
Attribution remains open, and this point calls for candor. Artex is a Chinese open-source tool, and the targets include China’s geopolitical rivals. But a Chinese open-source tool, operated from addresses spread across more than 10 countries, does not amount to a proven Chinese state operation. None of the sources published so far supports that bridge; the Japanese government speaks of scenarios under investigation, including foreign ones, and the suspect in the South Korea attacks remains unidentified.
The question the case opens is how to defend networks when the attacker scales without hiring people. GMO showed the method, the government named the category and the victims are tallying the damage. The answer lies less in pointing flags and more in treating autonomous agents as a class of cyber weapon of their own, with detection, logging and response designed for autonomous systems.